Research

Systematic investigation → evidence-based analysis → authoritative recommendations.

<when_to_use>

• Technology evaluation and comparison

• Documentation discovery and troubleshooting

• Best practices and industry standards research

• Implementation guidance with authoritative sources

NOT for: quick lookups, well-known patterns, time-critical debugging without investigation phase

</when_to_use>

Track with TodoWrite. Phases advance only, never regress.

Phase Trigger activeForm

Analyze Request Session start "Analyzing research request"

Discover Sources Criteria defined "Discovering sources"

Gather Information Sources identified "Gathering information"

Synthesize Findings Information gathered "Synthesizing findings"

Compile Report Synthesis complete "Compiling report"

Workflow:

• Start: Create "Analyze Request" as in_progress

• Transition: Mark current completed , add next in_progress

• Simple queries: Skip directly to "Gather Information" if unambiguous

• Gaps during synthesis: Add new "Gather Information" task

• Early termination: Skip to "Compile Report" with caveats

Five-phase systematic approach:

1. Question Phase — Define scope

• Decision to be made?

• Evaluation parameters? (performance, maintainability, security, adoption)

• Constraints? (timeline, expertise, infrastructure)

2. Discovery Phase — Multi-source retrieval

Use Case Primary Secondary Tertiary

Official docs context7 octocode firecrawl

Troubleshooting octocode issues firecrawl community context7 guides

Code examples octocode repos firecrawl tutorials context7 examples

Technology eval Parallel all Cross-reference Validate

3. Evaluation Phase — Analyze against criteria

Criterion Metrics

Performance Benchmarks, latency, throughput, memory

Maintainability Code complexity, docs quality, community activity

Security CVEs, audits, compliance

Adoption Downloads, production usage, industry patterns

4. Comparison Phase — Systematic tradeoff analysis

For each option: Strengths → Weaknesses → Best fit → Deal breakers

5. Recommendation Phase — Clear guidance with rationale

Primary recommendation → Alternatives → Implementation steps → Limitations

Three MCP servers for multi-source research:

Tool Best For Key Functions

context7 Official docs, API refs resolve-library-id , get-library-docs

octocode Code examples, issues packageSearch , githubSearchCode , githubSearchIssues

firecrawl Tutorials, benchmarks search , scrape , map

Execution patterns:

• Parallel: Run independent queries simultaneously for speed

• Fallback: context7 → octocode → firecrawl if primary fails

• Progressive: Start broad, narrow based on findings

See tool-selection.md for detailed usage.

<discovery_patterns>

Common research workflows:

Scenario Approach

Library Installation Package search → Official docs → Installation guide

Error Resolution Parse error → Search issues → Official troubleshooting → Community solutions

API Exploration Documentation ID → API reference → Real usage examples

Technology Comparison Parallel all sources → Cross-reference → Build matrix → Recommend

See discovery-patterns.md for detailed workflows.

</discovery_patterns>

<findings_format>

Two output modes:

Evaluation Mode (recommendations):

Finding: { assertion } Source: { authoritative source with link } Confidence: High/Medium/Low — { rationale }

Discovery Mode (gathering):

Found: { what was discovered } Source: { where from with link } Notes: { context or caveats }

</findings_format>

<response_structure>

Research Summary

Brief overview — what investigated, sources consulted.

Options Discovered

1. Option A — description
2. Option B — description

Comparison Matrix

Recommendation

Primary: [Option Name]

Rationale: reasoning + evidence Confidence: level + explanation

Alternatives

When to choose differently.

Implementation Guidance

Next steps, common pitfalls, validation.

Sources

• Official, benchmarks, case studies, community

</response_structure>

Always include:

• Direct citations with links

• Confidence levels and limitations

• Context about when recommendations may not apply

Always validate:

• Version is latest stable

• Documentation matches user context

• Critical info cross-referenced

• Code examples complete and runnable

Proactively flag:

• Deprecated approaches with modern alternatives

• Missing prerequisites

• Common pitfalls and gotchas

• Related tools in ecosystem

ALWAYS:

• Create "Analyze Request" todo at session start

• One phase in_progress at a time

• Use multi-source approach (context7, octocode, firecrawl)

• Provide direct citations with links

• Cross-reference critical information

• Include confidence levels and limitations

NEVER:

• Skip "Analyze Request" phase without defining scope

• Single-source when multi-source available

• Deliver recommendations without citations

• Include deprecated approaches without flagging

• Omit limitations and edge cases

Deep-dive documentation:

• source-hierarchy.md — authority evaluation details

• tool-selection.md — MCP server decision matrix

• discovery-patterns.md — detailed research workflows

Related resources:

• FORMATTING.md — formatting conventions